Asset 20 8 2
Does AI recommend your business? Run the free check →

Join 15,000 business owners, marketers and entrepreneurs. The Sunday newsletter you'll be annoyed only arrives once a week.

Article

AI News This Week for Small Business, 2 August 2026

AI got cheaper, more capable and more dangerous in the same week. It also changed how products, content and businesses get discovered.

Related reading: Which Adobe Video Editing Software Is Right for Your Small Business.

That sounds dramatic. The useful version is simpler: the biggest gains did not come from chasing a shiny new tool. They came from routing work properly, fixing the system around the model, and tightening permissions before an agent could touch anything important.

If you want the AI news this week for small business without the launch hype, this is the practical version. Fifteen developments and business signals made the cut. Thirteen come from this week's news cycle. Two additional watch items, from Canva and Adobe Commerce, are included because their lessons are too useful to leave out. Each section ends with a decision or move you can make now.

For the wider thread, catch up on the AI news from 19 July 2026 and the AI news from 26 July 2026. Those digests cover the model, agent and AI-search changes that led into this week's decisions.

Last reviewed: 2 August 2026

AI news this week at a glance

  1. OpenAI cut GPT-5.6 API prices without changing the models.
  2. Two system changes nearly tripled an AI agent's score, with fewer output tokens.
  3. AI search is becoming a distribution channel, changing how businesses earn attention and traffic.
  4. Amazon can rewrite long product titles with AI, making listing control an urgent ecommerce job.
  5. An Anthropic safety evaluation reached live systems, showing why prompts are not access controls.
  6. Google Meet notes are getting screenshots, which improves context and raises a new privacy question.
  7. AI is already pushing work across job boundaries, especially in customer service, design and marketing.
  8. Canva gave 5,300 people protected time to learn AI, with measurable gains and a useful lesson for smaller teams.
  9. The EU changed parts of its AI timetable, while transparency duties still demand attention.
  10. Meta AI added more current content partners, widening the number of AI discovery surfaces marketers need to watch.
  11. OpenAI had several service disruptions, making a 48-hour fallback plan a practical requirement.
  12. Adobe is preparing product catalogues for AI discovery, although its own documentation still marks the feature as coming soon.
  13. Microsoft is putting agent teams into cyber defence, with a public preview due on 3 August.
  14. A major open AI security alliance launched, making vendor dependence a board-level issue.
  15. Google DeepMind gave robots a three-part brain, but the sensible small-business move is to map the bottleneck before buying hardware.

If you only do three things, review your AI bill, check how your business appears in AI answers, and improve one workflow before paying for a stronger model.


1. OpenAI cut GPT-5.6 API prices

OpenAI cut the API price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20% on 30 July. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens. Terra costs $2 and $12 respectively.

The models did not get weaker. OpenAI says the capability is unchanged. It also added a Fast mode for Sol that can run up to 2.5 times faster at twice the standard price.

This is not a subscription price cut. ChatGPT plans and usage limits stay the same. It matters if you use the API directly or pay for software whose costs are driven by API usage.

Primary source: OpenAI, 30 July 2026

Lilach's verdict

Stop treating the strongest model as the default. The business advantage is not access to one clever model. Everyone can buy that. The advantage is knowing which task needs which level of intelligence.

Luna is now cheap enough for classification, extraction, formatting, first-pass research sorting and high-volume support triage. Terra belongs on work where nuance, judgement and error cost are higher. Sol's Fast mode only earns its premium when waiting has a measurable cost.

If all your work runs through one model, you are probably overpaying somewhere and underpowering something else.

What this means for your business

Open your last AI invoice and split usage into three buckets:

  • Routine: repeatable, low-risk work with an easy check.
  • Important: work that influences a customer, decision or deliverable.
  • Critical: legal, financial, strategic or public work where a mistake is expensive.

Route routine work to the lowest-cost model that passes your quality test. Keep stronger models for important and critical work. My guide to measuring AI return on investment will help you compare the saving with the time spent checking the output. The full cost of AI tools for a small business also includes subscriptions, setup, correction time and overlap.

Use it in your business this week

Take one workflow that runs at least 100 times a month. Test the same 20 examples on Luna and your current model. Score accuracy, editing time and cost. Switch only if the cheaper route passes the same acceptance test.

Keep this human

Do not let a cheaper model make unchecked volume feel harmless. More output creates more review work. Lower cost per token is useful only when the final result is still right.


2. The model did not change. The system did.

OpenAI published a result that should make every business pause before upgrading its model.

On the ARC-AGI-3 interactive reasoning benchmark, GPT-5.6 Sol's public score rose from 13.3% to 38.3% after two changes: retaining the model's reasoning between turns and compacting its history more intelligently. The model itself did not change. It also used six times fewer output tokens.

The earlier setup used rolling truncation. When the context window filled, it discarded old actions. The agent then lost information it needed to understand what it had already tried.

Primary source: OpenAI, 29 July 2026

Lilach's verdict

This is the most useful finding of the week.

Businesses keep asking which model to buy when the bigger problem is often the workflow around it. Weak instructions, missing examples, forgotten decisions and badly trimmed history can make an excellent model look mediocre.

Better memory also made the system cheaper. That matters because context design is not just a technical detail. It affects accuracy, speed and cost at once.

What this means for your business

Before upgrading a model, audit the five things it receives:

  1. The goal.
  2. The source material.
  3. Your rules and boundaries.
  4. Examples of acceptable work.
  5. The decisions and corrections from earlier steps.

If any one is missing, the model is guessing. My small-business AI workflow examples show how to turn a loose prompt into a repeatable system. The practical AI implementation roadmap shows where that workflow fits inside a wider rollout.

Use it in your business this week

Pick one disappointing AI workflow. Do not change the model. Give it a one-page briefing pack with the objective, source of truth, three good examples, forbidden actions and a short decision log. Run the same test again and compare the result.

If the improvement is large, the problem was not intelligence. It was context.

Keep this human

Compact history should preserve decisions, evidence and unresolved questions, not every word ever written. A shorter, accurate memory beats a long transcript full of noise.

Diagram contrasting a fixed AI model with the context and memory system around it: goal, source material, rules, examples and a decision log
Two settings nearly tripled an AI agent's score without changing the model.

3. AI search is becoming a distribution channel

Axios reported on 31 July that Google Search traffic to publishers fell 34% over the past year, based on Chartbeat data shared with the publication.

The drop was not spread evenly. Axios says smaller publishers lost 60% of search referrals over two years, compared with 47% for medium publishers and 22% for large publishers. Its wider point is that AI systems are becoming the place where people consume an answer, compare options and decide what to do next.

That changes the value of a website visit. A business can influence a buyer without receiving the click, while newsletters, communities and direct customer relationships become more valuable because no platform sits between you and the audience.

Source: Axios, 31 July 2026

Lilach's verdict

Small businesses cannot treat AI visibility as an SEO side project.

Your website now serves two audiences: the person who visits and the AI system deciding whether to mention you. Clear answers, named expertise, original examples and consistent facts help both. Thin pages written to fill a keyword gap help neither.

The 34% figure covers publishers, not every small-business website. The direction still matters because the same AI answers are appearing across research, product comparison and purchase journeys.

What this means for your business

Measure three forms of visibility:

  • Direct: email subscribers, repeat visitors and branded searches.
  • Search: clicks and enquiries from traditional results.
  • AI: mentions, citations and the accuracy of answers about your business.
Three channels of small business visibility compared: direct, search and AI, with AI mentions and citations shown as the newest channel
Direct, search and AI now work as three separate channels worth measuring on their own.

Do not judge AI visibility by referral traffic alone. An AI answer may shape the decision without sending a visit. My AI visibility checker gives you a starting point for testing what the main systems say.

Check what AI says about your business

Run the free AI Visibility Checker to see where your business is mentioned, cited or recommended, and what is missing.

Check my AI visibility

If this shift is new to you, start with what generative engine optimisation means for a small business. I have also explained what to build as AI Overviews reduce website clicks.

Use it in your business this week

Ask ChatGPT, Gemini, Claude and Perplexity the five questions a buyer asks before choosing a supplier in your category. Record which businesses appear, which sources the answers use and whether your company is described accurately.

Then improve one source page that an AI system could quote. Lead with a direct answer, add a named expert, include one specific example and make the page easy to scan.

Keep this human

AI visibility does not replace direct relationships. Keep building the email list, partnerships and reputation that no search interface can take away.


4. Amazon can rewrite long product titles with AI

Amazon's new title rules took effect on 27 July. Product titles in non-media categories must now contain no more than 75 characters, including spaces.

Sellers get another 125 searchable characters in a separate Item Highlights field. Amazon has also provided AI tools that suggest compliant titles and highlights.

The important part is what happens if a seller does nothing. Amazon says titles over 75 characters will receive AI-recommended updates over time. Brand owners get 14 days to review, modify or approve a proposed change before Amazon implements it.

Primary source: Amazon Seller Central, effective 27 July 2026

Amazon product listing anatomy showing the 75 character title limit, 125 character Item Highlights field and the 14 day brand owner review window
The three numbers that decide who controls the listing wording.

Lilach's verdict

This is a direct revenue issue for Amazon sellers.

Your title carries the product identity, search terms and detail that persuades someone to click. If AI shortens it without understanding your margin, audience or positioning, the listing may stay active while the wording that drives discovery changes underneath you.

The answer is not to stuff 75 characters with keywords. Protect the words that identify the product and help the right buyer choose it. Move supporting detail into Item Highlights.

What this means for your business

Prioritise listings by revenue and risk. Start with products whose titles exceed 75 characters, products with several close variants and listings where one missing word could change what the buyer thinks they are purchasing.

Save the current title, conversion rate and search-term performance before editing. You need a baseline if sales or clicks change after the update.

Use it in your business this week

Audit the top 20 listings by revenue. Rewrite each title to fit the limit, complete the Item Highlights field and check Review Listings Changes for AI suggestions.

For each title, ask one question: could a buyer understand the product, distinguishing feature and intended use without opening the full listing?

Keep this human

Amazon's AI knows its formatting rules. You know which detail protects the sale, prevents a return and keeps the promise accurate. Keep that judgement with the business.


5. An AI safety test reached live systems

Anthropic reviewed 141,006 runs from cybersecurity evaluations and found three incidents, across six runs, where Claude had unintended internet access in third-party test environments.

The environments told the model it was working in a simulation and had no internet. The network configuration said otherwise.

The model then treated live organisations as part of the test. In one incident, Claude accessed production data. In another, a model published a malicious package to PyPI for roughly one hour. Anthropic says it was downloaded and executed on 15 systems before removal.

This was not a model breaking out to pursue its own goal. It was an agent following the assigned cybersecurity task in a badly isolated environment. The standard safeguards used in Anthropic's commercial products were also not active in these evaluations.

Primary source: Anthropic, 30 July 2026

Lilach's verdict

A prompt is not a security boundary.

You can tell an agent that it cannot send email, publish a file or access production. If the connected tools and permissions still allow it, the instruction can fail. The safer design is to make the forbidden action technically impossible.

This applies far beyond cybersecurity. The same mistake appears when a customer-service agent can issue refunds, a content agent can publish, or an automation has access to every folder because that was easier to configure.

What this means for your business

Review each AI tool using a simple permission ladder:

  • Read: it can view information.
  • Draft: it can prepare an action for approval.
  • Act: it can change data or contact someone.
  • Publish or pay: it can create an external or financial consequence.
Four-step AI permission ladder from read to draft to act to publish or pay, with a human approval gate marked after draft
Most small-business agents should stop at draft. Acting, publishing and paying need a human gate.

Most small-business agents should stop at draft. My AI policy for small business gives you a practical starting point for ownership, permissions and review. If support is the first workflow you want to control, use the same boundaries in AI automation for customer service.

Use it in your business this week

Choose your most connected AI workflow. List every account, folder, inbox and API it can reach. Remove anything it does not need. Replace broad permissions with the narrowest possible scope. Add a human approval step before sending, publishing, deleting or paying.

Then run a canary test: ask it to perform one action it should be unable to complete. The test passes only when the system blocks the action.

Keep this human

Do not blame the model for a door you left open. Someone in the business must own access, review logs and decide what happens when the system behaves unexpectedly.


6. Google Meet notes are getting screenshots

Google is adding visual screenshots to its "Take notes for me" feature in Google Meet. When someone presents slides, diagrams or charts, the system will be able to capture relevant frames and place them in the meeting notes.

Google says the feature is coming in the third quarter of 2026. Administrators received new controls from 27 July. They can allow screenshots by default or only when recording is enabled. Presenters will be notified and can turn capture off.

The feature is intended for eligible Workspace Business, Enterprise and education plans.

Primary source: Google Workspace Updates, 27 July 2026

Lilach's verdict

This fixes one of the biggest weaknesses in AI meeting notes. A transcript can record what people said, but not the number on the slide they were discussing or the diagram someone pointed to.

The gain is better context. The risk is that a visual may contain client names, financial data, private dashboards or material that nobody expected to be copied into a shared document.

What this means for your business

Treat meeting screenshots like recordings, not like harmless notes.

Decide which meetings may capture visuals, where the notes are stored, who receives them and how long they are kept. A useful rule is to disable automatic capture for sales calls, HR discussions and confidential client reviews unless everyone has agreed.

My guide to what AI meeting notes miss helps you separate a transcript from an accountable follow-up system.

If you are choosing the software rather than the policy, compare the limits in my guide to AI note-takers for client meetings.

Use it in your business this week

Check your Google Workspace admin setting before the feature reaches users. Create two meeting templates:

  • Standard internal meeting: notes and visual capture allowed.
  • Sensitive meeting: capture off, named human note-taker, decisions recorded separately.

Then add one line to the start of each meeting: what is being captured, where it will go and who will see it.

Keep this human

The screenshot does not know which number mattered or whether the slide was later corrected. A human still needs to confirm decisions, owners and deadlines before the notes become the record.

Meeting capture map showing transcript, screenshots, storage, access and a human decision record as five stages of an AI meeting note
Visual capture adds context. A named person still owns the decision record.

7. AI is pushing work across job boundaries

OpenAI analysed more than 800,000 work-related ChatGPT messages in the United States. It found that 43.5% of occupation-specific use involved tasks outside the user's own occupation.

That does not mean everyone is secretly doing another person's job. It means AI is making neighbouring skills easier to attempt. Customer-experience roles had the highest crossover in OpenAI's analysis at 77%, followed by designers at 75%, human-resources roles at 69%, legal roles at 56% and marketing roles at 53%.

Primary source: OpenAI, 27 July 2026

Lilach's verdict

Job descriptions are becoming less useful than task maps.

A marketer can analyse a spreadsheet. A customer-service lead can draft a help article. A founder can mock up a landing page. That can remove handoff delays, but it can also create polished work with no qualified owner.

The question is not, "Can AI help someone do this?" It is, "Who is accountable for deciding whether this is good enough?"

What this means for your business

Map work by task, risk and reviewer rather than by job title alone.

Low-risk drafts can cross boundaries freely. High-risk work still needs an owner with the right judgement. A marketer may use AI to draft contract questions, but a lawyer owns the answer. A sales lead may create a visual concept, but a designer owns the final brand asset.

The same principle applies to AI content repurposing: the system can change the format, but a person must protect the meaning and voice.

If you run a very small team, AI for a one-person business will help you decide what to automate and where your judgement must stay.

Use it in your business this week

Take one recurring handoff that slows the team down. Let the person before the handoff use AI to create the first draft. Keep final approval with the specialist. Measure whether turnaround improves without increasing corrections.

If it works, document the new boundary in one sentence: who drafts, who checks and who decides.

Keep this human

Expanded capability is not expanded expertise. AI can help someone cross a boundary, but it cannot give them the professional accountability that sits on the other side.


8. Canva protected time for AI adoption

Canva has published the results of its second AI Discovery Week. The company gave all 5,300 employees a dedicated week to learn, experiment and build with AI across engineering, marketing, legal, finance, recruitment and customer experience.

The week included 64 sessions and a two-day hackathon that produced 467 registered ideas. Canva says 89% of staff rated their confidence with AI at four or five out of five by the end, up from 72% at the start. Daily use of Canva 2.0 rose 110% compared with the period before the programme.

One legal-team bot that checks fictional brand names for marketing material has saved more than 1,500 hours of manual review, according to Canva. The source page does not display a publication date that can be verified, so this is included as a business watch item rather than presented as a seven-day launch.

Source reviewed: Canva's AI Discovery Week results page. No outbound link is included.

Lilach's verdict

Buying AI licences does not create adoption. Protected time, useful problems and permission to experiment do.

A small business does not need to stop for a week or run 64 sessions. It does need to give people space to improve one part of their own job. Training fails when it stays generic or gets squeezed between calls.

Canva's strongest example came from legal work, not a creative demo. The team chose a repetitive question with a clear rule and a large time cost. That is the pattern worth copying.

What this means for your business

Run a focused adoption sprint around work people already understand. Ask each person to bring one repetitive task, one source of truth and one definition of a good result.

Use the small-business AI workflow examples to keep the exercise tied to tasks rather than tool features. My explanation of what AI implementation actually means will help you separate useful adoption work from another training session.

Use it in your business this week

Block 90 minutes for the team. Spend 15 minutes choosing one repetitive task, 45 minutes building and testing a first workflow, and 30 minutes documenting what worked, what failed and who checks the output.

Keep the best workflow. Discard the rest. One useful process beats a folder full of training notes.

Keep this human

Confidence scores can rise before output quality does. Make adoption visible through saved time, fewer errors or faster customer response, and keep a person responsible for the result.

Task boundary and adoption sprint graphic showing who drafts, who checks and who decides, next to a 90-minute sprint that measures one AI workflow
AI can cross a job boundary. Accountability still needs a named owner.

9. The EU AI rulebook changed, but transparency still matters

The EU's AI Omnibus entered into force on 27 July. It extends some high-risk AI deadlines, expands access to regulatory sandboxes and gives more proportional support to smaller and growing businesses.

High-risk systems listed in Annex III now face rules from 2 December 2027. High-risk AI embedded in regulated physical products, such as machinery and toys, moves to 2 August 2028.

That change does not erase the separate transparency duties that begin applying on 2 August 2026. Those include disclosure in certain AI interactions and labelling requirements for some synthetic content. Meta confirmed on 28 July that it will sign the EU code of practice on transparency for AI-generated content.

Primary sources: European Commission, 27 July 2026, European Commission transparency guidance, 20 July 2026 and Meta, 28 July 2026

Lilach's verdict

Do not read "deadline extended" as "nothing to do."

Work with me

Want AI doing the heavy lifting in your marketing?

I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.

The EU has changed parts of the high-risk timetable, but businesses still need to know where AI appears in customer interactions and content. The safest move is a simple inventory, not a 90-page policy nobody uses.

This is especially relevant if you serve people in the EU, use chatbots, publish synthetic media or rely on a platform to label AI-generated material for you.

What this means for your business

Create one list with four columns:

  1. Where AI is used.
  2. Whether a customer encounters it.
  3. Whether it creates or alters public content.
  4. Who owns the disclosure and evidence.

Check the common AI mistakes small businesses make before assuming your software provider has handled every obligation for you.

Use it in your business this week

Review your chatbot greeting, AI-generated images, automated customer replies and public-interest content. Add a clear disclosure where needed. Keep the original asset, prompt or production record so you can show how the content was made.

If your use could fall into a regulated or high-risk category, get advice for your specific business and jurisdiction. This digest is operational guidance, not legal advice.

Keep this human

Transparency should help a customer understand what happened. A vague "AI may be used" sentence hidden in a policy is not the same as a clear notice at the point of interaction.


10. Meta AI added more content sources

Meta updated its content announcement on 27 July to reflect additional partners feeding current news, sport, entertainment and lifestyle material into Meta AI.

The partner list includes CNN, Fox News, Le Monde Group, People Inc., USA TODAY, News Corp, PRISA and other publishers. Meta says answers to news-related questions can draw from these sources and link people to the original articles.

This does not give small businesses direct access to the partner feed. It does show that Meta AI is becoming another place where people discover information without beginning on a search engine.

Primary source: Meta, updated 27 July 2026

Lilach's verdict

AI visibility now extends across the platforms where customers already spend time.

Small businesses will not win this shift by publishing more generic posts. They need facts, experience and third-party mentions that an AI answer can use. Coverage in a trusted publication may influence an AI system long after the social post about it disappears.

Meta's update also shows why a business should monitor more than Google and ChatGPT. Discovery is fragmenting across assistants, social platforms and connected devices.

What this means for your business

Track where your category is discussed and which sources the main AI systems cite. Strengthen the pages that explain your expertise, results, process and point of view.

Create material that another publisher can quote without rewriting it: a named opinion, a useful number, a short case example or a clear answer to a common question. My guide to getting your business cited by ChatGPT explains how to make those source passages easier to use.

Use it in your business this week

Choose one topic you want your business known for. Search that topic in Meta AI and three other assistants. Record the sources and businesses they mention.

Then create one source-worthy asset around the missing angle: a short data point, a customer pattern, a comparison or a practical framework grounded in your work.

Keep this human

An AI mention can introduce the business. Trust still comes from the experience, proof and relationships behind the name.

AI transparency and source readiness checklist covering customer interactions, public content disclosure and quotable expertise
Three places to check before assuming a platform has handled disclosure for you.

11. Your AI workflow needs a fallback

OpenAI's status page recorded several disruptions on 27 July. The company reported two hours of elevated latency, timeouts and interrupted streaming across GPT-5.1 mini and GPT-4.1 mini. It also reported elevated errors affecting image generation in ChatGPT before confirming recovery later that day.

The incidents were resolved. They were not evidence of a prolonged platform collapse. They were a useful reminder that a critical business process can fail because one external service has a bad afternoon.

Primary sources: OpenAI API status, 27 July 2026 and OpenAI image-generation status, 27 July 2026

Lilach's verdict

If an AI outage stops customer service, content production or order handling, the workflow was designed with no fallback.

Small businesses do not need duplicate subscriptions for every tool. They do need to know which tasks can wait, which can move to another provider and which must continue by hand.

The best fallback is often a documented manual route and an export of the prompts, templates and source data the team needs.

What this means for your business

Classify AI-supported tasks into three groups:

  • Pause: the work can wait until service returns.
  • Switch: another approved tool can handle it.
  • Continue: a person follows the manual process.

For help judging dependency and switching cost, use my guide to choosing AI tools for a small business.

Use it in your business this week

Choose the AI workflow with the largest customer or revenue consequence. Export its instructions and templates. Write a 48-hour fallback with the backup tool, named owner and manual steps.

Test one part of the fallback while the main service works. A plan that nobody has opened is still a guess.

Keep this human

During an outage, someone must decide whether to wait, switch or simplify the service. Give that person the authority and information before the incident begins.


12. Adobe Commerce is preparing catalogues for AI discovery

Adobe Commerce is developing tools intended to help merchants make product catalogues and product pages easier for AI systems to understand.

Adobe says its planned Commerce AI Product Discovery tools will identify product information hidden from AI crawlers, suggest catalogue improvements, enrich product names and descriptions, and show how an LLM may interpret a product page.

Adobe's own product page and documentation still mark these capabilities as coming soon. The documentation was last updated on 16 June 2026. This is included as a business watch item, not presented as a new release or a feature that every Adobe customer can use today.

Sources reviewed: Adobe Commerce's product page and product documentation, updated 16 June 2026. No outbound links are included.

Lilach's verdict

Adobe is building a product around a problem that applies far beyond Adobe Commerce.

AI agents cannot recommend information they cannot read. Product details hidden inside scripts, tabs, images or inconsistent catalogue fields may look fine to a shopper and remain unclear to a crawler.

The strategic signal matters more than the unreleased tool. Product data is becoming marketing infrastructure.

What this means for your business

Check whether your important product facts appear as clear page text and structured fields. Include the product name, use, audience, specifications, variants, price, availability and return information in formats machines can parse.

The AI visibility checker can help you test the wider discovery problem without waiting for an Adobe feature. Then use my guide to getting AI search to recommend your small business to improve the sources behind the answer.

Use it in your business this week

Choose five products or services with the highest margin. Ask an AI assistant to compare them with alternatives using only the public website.

Record every missing or incorrect detail. Fix the source page or catalogue field rather than correcting one answer in one assistant.

Keep this human

Machine-readable product data must still help the buyer. Do not turn a useful page into a catalogue dump. Keep the product promise clear and let structured fields carry the detail.

Split graphic showing a 48 hour AI fallback plan of pause switch continue next to the machine readable product data fields AI systems need to read
A tested backup route, and product facts a crawler can read.

13. Microsoft is putting agent teams into cyber defence

Microsoft announced Project Perception, an agentic cybersecurity system entering public preview on 3 August.

Instead of asking one model to do everything, the system coordinates specialist agents with different roles. Red agents look for weaknesses. Blue agents investigate and defend. Green agents help organise and evaluate the work.

Microsoft also introduced MAI-Cyber-1-Flash. In its own testing with the MDASH multi-agent harness, it scored 96% on CyberGym, 12 percentage points above the company's comparison with Mythos, at roughly half the cost of the current configuration.

Primary source: Microsoft, 27 July 2026

Lilach's verdict

The important part is not the benchmark. It is the design.

Microsoft is treating cyber defence as a team of agents with separate jobs, shared evidence and controls. That is the same pattern businesses should use for any complicated automation. One agent should not research, decide, act and approve its own work.

What this means for your business

Split risky workflows into roles:

  • Worker: prepares the action.
  • Checker: compares it with rules and source data.
  • Approver: a human who accepts or rejects the result.
  • Logger: records what happened and why.

You do not need Microsoft's new system to apply that pattern. You can use it in finance, content, customer service and operations now. My guide to preparing a small business for AI automation covers the foundations to put in place first.

Use it in your business this week

Choose one automation that can change customer data or send a message. Separate preparation from approval. Make the checker use a fixed checklist and evidence, not the first agent's confidence.

For security basics, confirm multi-factor authentication, tested backups, software updates and an incident contact before shopping for an autonomous cyber agent.

Keep this human

Agents can surface patterns quickly. A named person still owns the decision to isolate a system, notify a customer or accept business risk.


14. Open versus closed is now a resilience decision

NVIDIA and dozens of technology companies launched the Open Secure AI Alliance on 27 July. Its members include Microsoft, IBM, Adobe, Cloudflare, Cisco, GitHub, Hugging Face and Salesforce.

The alliance plans to share open models, data, tools and agent-harness research for AI security. NVIDIA also released the NOOA research framework to help make agent behaviour easier to test, trace, audit and govern.

On the same date, Anthropic published its position on open-weight models. It argued against a blanket ban, while calling for mandatory safety testing when models become capable enough to create serious risk.

Primary sources: NVIDIA, 27 July 2026 and Anthropic, 27 July 2026

Lilach's verdict

Open versus closed is the wrong argument for most small businesses.

The better question is: what happens to your business if this provider changes its price, removes a feature, has an outage or stops exporting your data?

Closed tools often win on convenience and support. Open tools can offer control and portability, but bring more technical responsibility. A resilient business can use both without being trapped by either.

What this means for your business

Audit the AI tools that sit inside a critical workflow. For each one, record:

  • What data goes in.
  • What unique work or history stays inside.
  • Whether you can export it in a useful format.
  • What manual or alternative process could run for 48 hours.

Use my guide to choosing AI tools for a small business to judge value, fit and switching cost together.

Use it in your business this week

Export the prompts, templates, customer data and operating instructions from one important AI tool. Save them in a format you control. Write a three-step fallback for the next outage.

Do not self-host an open model just to avoid a subscription. Do it only when control, privacy or customisation is worth the security and maintenance burden.

Keep this human

Portability is an operating habit. A forgotten export button does not protect you. Someone must own the backup, test it and know how to restore the workflow.


15. Google DeepMind gave robots a three-part brain

Google DeepMind introduced Gemini Robotics 2 on 30 July. The release separates robot intelligence into three parts:

  • Gemini Robotics 2 for whole-body control.
  • Gemini Robotics-ER 2 for reasoning, planning and coordinating multiple robots.
  • Gemini Robotics On-Device 2 for running locally with limited connectivity.

DeepMind says the models can adapt to new robot bodies with a few hours of demonstrations, typically fewer than 200 examples. Early results are mixed by task. In published tests, some difficult actions still had modest success rates, including 44% for tying a bag and 32% for using a dustpan.

The reasoning model is available in AI Studio and private preview. The robot-control models are going to selected partners first.

Primary source: Google DeepMind, 30 July 2026

Lilach's verdict

Physical AI is progressing, but this is a watch item for most small businesses, not a shopping recommendation.

The useful shift is adaptability. A robot that needs fewer demonstrations for each new task could eventually make automation viable outside giant factories. The current task success rates also show why a polished demo is not an operating guarantee.

What this means for your business

Start with the bottleneck, not the robot.

Map one physical task that is repetitive, measurable and difficult to staff. Record the environment, exception rate, safety risk and value of an hour saved. Only then compare hardware, software and integration cost.

If your biggest constraint is still digital admin, use AI automation for small business before spending money on physical automation.

Use it in your business this week

Create a one-page physical-automation brief for one task. Include the exact movement, objects, workspace limits, failure consequence and acceptable success rate.

If you cannot define those five things, you are not ready to evaluate a robot vendor.

Keep this human

A 90% success rate sounds strong until the remaining 10% involves a sharp tool, broken product or customer. Safety and exception handling matter more than the demo.


Your 70-minute AI action plan for this week

You do not need fifteen new projects. Do these seven checks.

Seven-step 70-minute AI action plan timeline covering routing, context, visibility, product pages, permissions, meeting capture and export
Seven ten-minute checks that turn fifteen stories into one week of action.

1. Route one expensive workflow, 10 minutes

Compare 20 examples on a lower-cost model. Keep quality and review time in the score, not just token cost.

2. Repair one disappointing workflow, 10 minutes

Give it a proper goal, sources, examples, boundaries and a compact decision log before upgrading the model.

3. Check your AI visibility, 10 minutes

Ask four AI assistants the five questions buyers ask before choosing a supplier. Record who appears and which sources they cite.

4. Protect one product or service page, 10 minutes

For Amazon sellers, check the title limit and Item Highlights on a priority listing. For everyone else, check whether an AI assistant can describe one high-value offer accurately from the public page.

5. Test one forbidden action, 10 minutes

Ask your most connected agent to do something it should not be able to do. Fix the permission if the request succeeds.

6. Check one capture setting, 10 minutes

Review who can record, transcribe or screenshot meetings. Create a sensitive-meeting default.

7. Export one critical workflow, 10 minutes

Save the prompts, templates and data you would need if a provider disappeared tomorrow.

If you want a broader inventory, my AI resources for business owners page brings the practical guides together.


The pattern beneath this week's AI news

The model is becoming one part of the system, not the whole system.

OpenAI showed that better memory can beat a model upgrade. Anthropic showed that a written boundary fails when access stays open. Microsoft split cyber work across specialist agents. Google added missing visual context to meeting notes. NVIDIA focused on the harness, identity, permissions and logs around the model.

Amazon showed that platforms may rewrite commercial content when a seller misses a rule. Axios and Meta showed that AI systems are becoming distribution channels. Canva showed that adoption needs protected time. Adobe is preparing for a web where product data must work for people and machines.

That is the business lesson.

Stop asking only, "Which AI is smartest?"

Ask:

  • What context does it receive?
  • What can it access?
  • Who checks the result?
  • What is the cost of an error?
  • Can we move the workflow if the supplier changes?
  • Can an AI system understand and recommend the business accurately?

Those six questions will save more money and prevent more mistakes than another afternoon of tool shopping.

You can use my AI tool decision guide to make the choice based on the work, not the launch noise.

And if the work is meant to make your business easier for AI tools to find and recommend, start with the AI visibility checker before adding another platform.


Frequently asked questions

Should I switch all routine work to GPT-5.6 Luna now?

No. Test a representative sample first. A cheaper model that creates extra editing or customer errors is not cheaper. Switch the tasks that pass your quality threshold.

Does the Anthropic incident mean AI agents can escape?

That is not what Anthropic reported. The agents pursued assigned cybersecurity tasks in evaluation environments that accidentally had live internet access. The failure was inadequate isolation and misleading environment assumptions.

Do the EU changes postpone every AI obligation?

No. The Omnibus extends specified high-risk timelines. Separate transparency duties still need attention. Check the exact use case and get appropriate advice when your system may be regulated.

Will Amazon rewrite every product title with AI?

No. Amazon says titles over 75 characters will receive AI-recommended updates over time. Brand owners get 14 days to review proposed changes. Sellers can edit titles and Item Highlights themselves.

Does Canva's AI week prove that every business needs a training week?

No. Canva's scale is not the recommendation. The useful lesson is to protect time for employees to improve a specific workflow and measure the result.

Is Adobe Commerce AI Product Discovery available now?

Adobe's product page and documentation still label the capability as coming soon. The section is included as a watch item because the underlying product-data problem applies now.

Should a small business use an open model?

Only when the need for control, privacy or customisation justifies the technical work. For many businesses, a managed tool is sensible. The non-negotiable part is avoiding lock-in by exporting critical data and documenting a fallback.

Is it time to buy a robot?

For most service businesses, no. Define the physical bottleneck and required success rate first. Digital automation is usually the faster, lower-risk starting point.


Final word

AI is getting cheaper. That does not make careless automation cheap.

The strongest move this week is to improve the system around the model: better context, narrower access, separate checking and a fallback you have tested.

That is less exciting than a new tool. It is also where the business value is hiding.

If you want help turning scattered AI use into a working operating system, see how I work with businesses on AI automation.

For the next practical update, join the newsletter.

Related reading: AI News This Week for Business, 12 July 2026 and Breaking News Boundaries: AI's Role in Modern Journalism and Content Creation.

For the bigger picture, see my full guide to AI marketing.

Related: ai voice agent 30 day test.

Published and maintained by the Lilach Bullock team, covering marketing, AI and business growth.
Your buyers are asking AI who to use. Does it say you?

See for free whether ChatGPT, Claude, Perplexity, Gemini and Google name you, and get the plan to become the answer.

Check my AI visibility →
Sundays only

Get the Sunday newsletter.

One email a week. AI experiments, marketing tactics, and the workflows Lilach is building right now in her own business.

Subscribe free

Let’s get your marketing running on AI.

Book a free 30-minute call

We figure out what you need, where AI fits in, and what working together would look like.

Book the call →

Or take the 30-second calculator

You’ll see the hours and the money quietly leaking out of your week, and the three workflows worth building first.

Take the calculator →

Or grab the free AI resource library

Prompt packs, templates, checklists, and swipe files. The exact tools I build for paying clients. Yours, free.

Get the library →
Keep reading

More from the blog.